A benchmark for automated peak picking of protein NMR spectra

In drug discovery field, one of the major used techniques is Nuclear Magnetic Resonance spectroscopy (NMR). To date, most of the steps in NMR analysis have been automated. The remaining step, peak picking, is still performed manually, what is extremely time-consuming and constitutes limiting factor in NMR-based drug discovery. Peak picking is a process of recognition of 2D/3D patterns in NMR image. To help automation of peak picking we have prepared a considerable benchmark, which enables testing various approaches. We also propose a baseline framework based on object detection scheme, which is well studied in the field of computer vision, and addresses automation of peak picking. Finally, we present the results of empirical studies, where a few different methods were compared, and show that peak pickers based on our framework significantly outperform the remaining methods for complex spectra and behave comparably for average ones.